Ant Group presents four papers at IJCAI 2026, systematically exploring AI's adaptive capabilities in dynamic environments.

Author and source: Leiphone

No matter how sophisticated the algorithm, it ultimately must be tested in a dynamic environment.

IJCAI-ECAI 2026 will be held in Bremen, Germany, from August 15 to 21, 2026. As the conference approaches, AI Science & Technology Review is systematically scanning the accepted papers to identify emerging technical trends and industry directions.

InHuawei IJCAI 2026 Paper Review: Shifting from “Scale Density” to “Design Density”, we observed major companies showing restraint in the parameter race and a focus on computational efficiency.

The four papers selected from Ant Group point to another equally important but distinctly different technological trajectory: a systematic exploration of adaptability in dynamic environments.

The accuracy of algorithmic models continuously breaks records in laboratory settings, but performance degradation becomes almost inevitable once deployed in the real world. The reason is simple: the real world is never static.

Ant Group's business scenarios push this "non-static" nature to the extreme: over a billion users, more than 8,000 services, payment peaks and midnight lulls differing by several orders of magnitude, risk control models required to respond to black-hat strategy iterations on an hourly basis, and global deployment needing to adapt to vastly different financial infrastructures across countries.

In such an environment, static models with the “train once, deploy forever” approach are destined to fall into the trap of “marking the boat to find the sword.”

It is precisely for this reason that true intelligence lies in whether an algorithm can perceive environmental changes, proactively adjust its strategies, and seek optimal solutions in dynamic conditions. These four papers from Ant Group collectively address this core question: How can AI evolve from a “static solver” into a “dynamic adaptive agent”?

01 MSRGC-Net: Enabling Adaptive Data Density Distribution in Time Series Clustering

Time series clustering automatically groups massive sequences of behavior, serving as a fundamental tool for understanding behavioral patterns and monitoring anomalous signals.

But existing methods each have their own drawbacks:

• Similarity methods (such as k-Shape): Can capture local patterns, but require computing pairwise distances between all samples, making them slow with large datasets;
• Deep learning methods (such as autoencoders): strong representation capabilities, but expensive to train, tedious to tune, and highly computationally intensive;
• Traditional feature extraction: Relies on manually designed features, potentially losing key temporal dynamic information.

The bigger issue is that real-world time-series data density is extremely uneven—transaction data during Double Eleven is as dense as a downpour, while data at 3 a.m. is as sparse as a single drip—but traditional methods require you to predefine how many categories to split into, which is inherently misaligned with reality.

MSRGC-Net, proposed by Ant Group in collaboration with Chongqing University of Posts and Telecommunications, avoids the trade-off between accuracy and computational efficiency with a set of "training-free" techniques.

Paper URL: https://arxiv.org/pdf/2606.12077v1

▎Its core logic can be broken down into three steps:

Step 1: Multi-scale reservoir encoding (feature extraction). Use multiple untrained echo state networks (ESNs) as the fundamental units for reservoir computing; by simply adjusting the spectral radius, local fluctuations and long-term trends from the original time series are simultaneously extracted. All samples are aggregated to form a view feature matrix, which serves as input for subsequent stages.

Step two: Constructing the granular ball anchoring map (structural modeling), the most ingenious part. Instead of focusing on individual data points, the algorithm automatically divides the data into "granular balls" based on local density: high-density regions form small, numerous granules, while low-density regions form large, fewer granules. The data scale is compressed from N to M (number of granules, M < N), while noise interference is also suppressed.

Step 3: Consensus-based multiscale graph optimization. Anchored graphs across scales are fused through a consensus strategy, enabling the model to capture both micro-level local patterns and macro-level global trends, ultimately generating robust clustering results without requiring manual specification of the number of clusters.

How effective is this strategy?

On five multi-variable benchmark datasets with 15 evaluation metrics (five each for NMI, ARI, and RI), MSRGC-Net achieved the best results in 12 metrics and second place in 2—an 80% best-rate.

More importantly, it avoids O(n²) pairwise computations, achieving near-linear complexity. In simple terms: it’s fast, accurate, and doesn’t require you to guess how many categories to split into.

In Ant's actual business scenarios, this means the system can automatically adjust clustering granularity based on traffic density—fine-grained clustering to detect anomalies during peak times, and coarse-grained summarization to save computing power during low periods, always following the data.

02 Enable adaptive data mixing strategies for offline-to-online reinforcement learning

Reinforcement learning has a well-known problem of "poor adaptability"—"distribution shift": a model trained on offline data tends to "forget" immediately upon deployment. This occurs because there is a distribution difference between offline data and online interaction data, causing the online policy to rapidly lose the knowledge learned from offline training.

Current solutions essentially fall into two categories: either maintaining a fixed mix ratio, such as 50% offline and 50% online, or using heuristic rules to prioritize sampling “on-policy” samples.

But the problem is that the requirements for the strategy are completely different at different stages—early on, you need more offline data to stabilize the foundation; later, you need more online data to explore new possibilities. A fixed strategy is like carving a mark on a boat to find a lost sword.

Ant Group, in collaboration with Shanghai Jiao Tong University, proposed ROAD, transforming the data mixing ratio from a "preset parameter" into a "dynamic decision variable."

Paper URL: https://arxiv.org/pdf/2605.14497

Its core idea is interesting: data selection is essentially a bilevel optimization problem.

• Inner-Level: Standard Q-learning update, minimizing the Bellman error;
• Outer-Level: Treat the data mixing strategy as a meta-decision variable to maximize the expected return of the online strategy.

The two layers are approximately solved using the Multi-Armed Bandit (MAB) algorithm to enable the hybrid strategy to adjust in real time during training: when the model begins to "forget" knowledge learned during the offline phase, the sampling ratio of offline data is automatically increased to maintain stability; when the model becomes overly conservative and lacks exploration, the proportion of online data is promptly increased to encourage experimentation.

The experiments were conducted rigorously. The team validated their approach on three classic offline-to-online reinforcement learning benchmarks: AntMaze (a robot navigating a maze to reach a goal), MuJoCo (biomechanical robot motion control), and FrankaKitchen (a robotic arm performing complex tasks in a kitchen). These tasks vary in difficulty and state space, providing a comprehensive evaluation of the algorithm’s generalization capability.

To verify the generalizability of ROAD, the research team integrated ROAD with various leading offline algorithms, including IQL, PEX, CQL, and Cal-QL. The results showed that, regardless of the underlying algorithm, ROAD consistently delivered performance improvements.

Taking PEX+ROAD as an example, this method achieves an overall average score of 71.12 across 24 continuous control tasks in the D4RL benchmark, ranking first in 18 of the 24 tasks, and significantly outperforming all baselines including fixed ratio, decreasing ratio, and heuristic balanced replay.

In other words, ROAD enables the model to find the optimal balance between “conservative inheritance” and “aggressive exploration”—avoiding failures at launch while not missing out on improvements from real-time data. This approach offers clear value in scenarios such as deploying Ant Group’s risk control models or optimizing recommendation systems in real time.

03 DSEBO: Enabling Adaptive Search Subspaces for High-Dimensional Bayesian Optimization

Bayesian optimization is a classic method for black-box function optimization, but it struggles with high-dimensional problems. A common solution is random embedding—projecting the optimization into a lower-dimensional subspace—but this raises a new question: what is the effective dimensionality?

Traditional methods either rely on expert experience to set fixed subspace dimensions or use trial-and-error estimation, consuming significant resources and failing to adapt fixed dimensions to different tasks. Worse still, the effective dimensionality itself may change during optimization: lower dimensions may suffice during initial exploration, while higher dimensions with finer details may be needed for fine-tuning later.

DSEBO, proposed by Ant Group in collaboration with East China Normal University and Nanjing University, treats subspace dimension as a dynamic variable that autonomously switches during the optimization process.

Paper URL: https://arxiv.org/pdf/2605.23473

The core mechanism of DSEBO is "ascending step by step from low to high":

• Starting from a low-dimensional subspace, the optimizer generates candidate solutions in the low-dimensional space, maps them randomly into the original space, quickly discerns the general shape of the objective function, and uses the resulting feedback to iteratively update the Gaussian process.
• After convergence is observed, dimension expansion is automatically triggered, inheriting the low-dimensional solution through a shared embedding matrix;
• Initialize and continue optimizing new subspaces, with a shared embedding matrix ensuring that all subspaces are nested within each other, forming a cycle of "low-dimensional exploration → convergence trigger → dimension expansion → continued optimization" until the evaluation budget limit is reached.

During the dimension expansion phase, the extent of expansion is not fixed—the algorithm adaptively adjusts based on the current optimal value curve: it expands slowly when the curve is flat to avoid unnecessary overhead, and rapidly when the curve is steep to quickly capture high-dimensional gains.

The experimental results show that, when compared against more than a dozen leading methods such as REMBO, SIRBO, BAxUS, and TuRBO on synthetic functions (Levy, Griewank, Sphere, D=1000) and three real-world tasks (MSLR, Lasso-Hard, LIMO), DSEBO achieved the optimal solution in nearly all tasks—leading in both accuracy and convergence speed.

This “low-dimensional exploration, high-dimensional fine-tuning” approach is also highly practical in Ant’s daily R&D—hyperparameter optimization for credit models, threshold optimization for risk control strategies, and multi-variable experimental design for marketing campaigns no longer require experts to guess dimensions manually; the algorithms can now “learn as they go,” enabling automated and highly efficient operations.

04 VGA-BenchV2: Enabling Video Evaluation Benchmarks to Adapt to the Evolution of the AIGC Ecosystem

If the first three papers discuss algorithmic adaptability, VGA-BenchV2 demonstrates how evaluation infrastructure can "adapt" to the evolution of the AIGC technology ecosystem. It is the only one of the four papers focused on multimodal content understanding, reflecting Ant Group's strategic investments in non-financial domains such as digital life and content ecosystems.

AIGC has triggered a boom in video production, but a core issue remains: how can we systematically evaluate the quality of these generated videos?

Traditional evaluation metrics such as FVD (assess overall quality and temporal consistency) and CLIP Score (evaluate semantic consistency between generated images and text descriptions) primarily focus on whether the visuals are clear, the motion is smooth, and the text matches accurately. However, they are nearly powerless when it comes to perceptual qualities related to “aesthetics”—such as composition, lighting, color harmony, and more.

At CVPR 2026, Ant Group, in collaboration with the Beijing Film Academy and the Big Artificial Intelligence Institute (BIGAI), introduced VGA-Bench, which established the first three-dimensional framework for video aesthetic evaluation (aesthetic quality, aesthetic labels, and generation quality). Built upon 1,016 prompts, 12 generation models, and over 60,000 videos, the benchmark enabled AI to "appreciate" video content from a professional perspective for the first time.

Paper link: https://arxiv.org/pdf/2604.10127

But video generation models are evolving too quickly; the monthly iteration pace has made fixed benchmarks quickly “insufficient.” In this IJCAI 2026 paper, the team further introduces VGA-BenchV2.

Open source address:

https://huggingface.co/datasets/BestVictoryLab/VGA-Bench

▎There are three key advancements:

First, the volume of human annotations has increased. VGA-BenchV2 adds 36,000 new task-level human annotations, including 16,200 aesthetic quality annotations, 13,200 aesthetic label annotations, and 6,600 generation quality annotations—representing increases of 13.46x, 11.15x, and 1.55x over VGA-Bench, respectively. With more abundant “human preference” data, the evaluator aligns more accurately with human aesthetic judgment.

Second, the evaluator architecture has been upgraded. The team designed a hybrid architecture: VAQA-Net handles continuous aesthetic scoring, while VTag-Net and VGQA-Net perform tag recognition and quality assessment based on the Qwen large vision-language model. The integration of large models has significantly enhanced the evaluator’s understanding of complex visual semantics. Experimental results show that its assessments across multiple generative models align closely with human judgments.

Third, the closed loop from "evaluation" to "optimization" is fully enabled. This is the most groundbreaking design of VGA-BenchV2: it uses the learned aesthetic evaluator as a reward signal to directly fine-tune generative models via reinforcement learning. This means the evaluation benchmark is no longer just a "referee"—its scores can be directly converted into optimization signals, guiding generative models to continuously improve their aesthetic quality.

From VGA-Bench to VGA-BenchV2, the evaluation system is no longer a static metric but a living system that co-evolves with the generative ecosystem. For Ant Group’s use cases in content understanding, security review, and recommendation optimization, video quality assessment not only keeps pace with the evolution of the AIGC ecosystem but also drives upstream model improvements in return: closing the loop from “scoring” to “optimization.”

Closing: No matter how sophisticated the algorithm, it must ultimately be tested in a dynamic environment.

Four papers approach the topic from four directions—unsupervised learning, reinforcement learning, black-box optimization, and multimodal evaluation—collectively pointing to the same technical pathway: a paradigm shift from static to dynamic.

Experimental data also confirms the effectiveness of this approach: MSRGC-Net achieved 12 first-place and 2 second-place results across 15 metrics; ROAD outperformed existing policies in multiple offline-to-online reinforcement learning tasks; DSEBO delivered measurable improvements on thousand-dimensional optimization problems; and VGA-BenchV2 demonstrated high alignment with human judgments across multiple generative models.

Looking at Ant's overall strategy along this technological path reveals a clear "dual-wheel drive" structure: one wheel deeply investing in time series modeling, reinforcement learning, and optimization algorithms within core financial scenarios; the other wheel proactively building multimodal understanding and evaluation infrastructure for digital life and content ecosystems.

Ultimately, no matter how sophisticated the algorithm, it must be tested in a dynamic environment. The true value of these four papers may lie in the fact that they all grew out of real business contexts: Ant’s business scenarios are not merely a backdrop for the papers, but the source of the problems, the origin of the data, and the proving ground for their effectiveness.

And perhaps this is precisely what makes AI research in industry so unique—they’re not just “writing papers,” but seeking solutions that can withstand dynamic real-world changes.